Attractor — Context Research
Comparing models and exploring recurring conversation patterns.
What do conversations return to, and how much of that pattern comes from the model used to read them?
Jennifer Naomi Nguyen · exploratory research tooling · March 2026 – present
The question
Across long conversations, I keep returning to familiar topics and ways of thinking. Attractor turns that history into a weighted map of engagement modes. I built the tool to carry useful context forward, then realized its output was also something I could study: do different models propose different maps from the same material, and can the supplied memory itself shape what appears later?
The tool and the research
The tool summarizes conversations and updates the map so a later session can receive a smaller, relevant context block. The research asks what that map means, whether the update depends on the model, and whether feeding the context back changes the pattern. Those two uses share one mechanism; the published visualization makes its current output inspectable.
Try the published view
Open the demo above to inspect the published, read-only state from this site's Cloudflare Worker. It shows basin labels, weights, keywords, and connections; opening it makes no model call and does not expose conversation transcripts. The largest basin, Provenance and documentation, remains visible because removing it would misrepresent the state.
A separate saved Luna/Astra comparison helped motivate the research, but it is not the live view shown here. In that one-run comparison, both models ranked research first while proposing different connections. A recorded Haiku/Opus comparison also showed different weight updates. These examples identify questions; they do not establish a stable model effect.
How the comparison works
Attractor summarizes a conversation and asks a model to update a starting map. Its compare command holds the conversation and starting state constant across model choices and saves the proposed summaries, weight changes, and connections. Because each model currently creates its own summary, a difference in the map cannot be attributed only to the update step.
Next I want to compare fixed summaries across repeated conversations, measure patterns independently of Attractor, and vary whether memory is supplied. Those controlled studies remain planned.